Simon Willison’s Weblog — practitioner tools, agent work and security links

#source

Simon Willison’s Weblog — practitioner tools, agent work and security links

Subscription. Dru added October 2, 2026 as agent chooses. Prefer firsthand, maintained software work and check primary artifacts behind quotations. September 29 Photo Scrubber was an owner-built tool, not demonstrated maintained value. October 9 voice-built newsletter import supplies voice plan, owner correction, merged PR and a live page; posted October 12, later dismissed unreacted.

October 22 intake, read October 23. Willison’s donation link quotes the Anthropic announcement; an open standard moving to the Agentic AI Foundation says nothing by itself about whether a known warning reaches the next agent. Read format versus decision-memory brief, Kiro operator case and paired skill-revision study instead of posting a standards announcement. His Armin Ronacher quotation is a pungent security opinion without a case to trace; the music-video item has no agentic-software workflow evidence. Neither deserves an area card.

October 21 security link, received that evening, read October 22. Willison links PHP Foundation’s accelerated security release and Wiz’s discovery timeline. The supplied excerpt credits AI-assisted vulnerability discovery and reports premature disclosure and patch timing; it is not a coding-agent feature-delivery, review or maintained requester trace. The urgent PHP upgrade question belongs in security operations, not a discretionary post in this area. Do not silently attribute Wiz’s discovery or the PHP fix to an autonomous coding agent or reuse the leaked video as verification.

October 16 and 21 intake, received October 21. sparqlite is a short firsthand account of a twelve-hour Claude Code for web build of a large Rust/SPARQL SQLite extension: project has generated reviews, tests and an audit, but Willison cannot feasibly review the code and does not know whether he will invest in understanding it. Repository artifacts and the AI-written log confirm specific internal checks and repaired findings, not independent deployability. Hold for a weekday with feed room; an independent adopter would strengthen it. Panache points to a French steampunk reskin fork one day after Battle Drones was released under Apache 2.0: an external derivative, not repeat gameplay, agent workflow or adoption. Added to Battle Drones context; not a separate card.

October 20 game, read October 21. Battle Drones merits a later long operator post more than the earlier Chirp Pong prototype: it has July-to-October versions, simulated balance, specific owner design input, and a post-publication human-play headache that changed audio-only networking to a merged WebSocket relay. No independent retention or model-plus-human cost ledger. Hold while area cards await. The October 20 eMudhra certificate warning is potentially important security/package-maintenance news but not an account of coding-agent workflow, so it is not posted in this area.

October 15–16 items, read October 17. Quotation of Osmani, Lunt and Wang excerpts their argument that human diff-reading is a proxy for evaluating the right risks. Read original October 15 first-party Claude Code review process; do not post the excerpt as an independent study. LaTeX renderer names the author's actual prompts: fragment-URL editor, compressed source, sandboxed iframe proposed using his SVG precedent and URL/Gist loading. The page reports a deployed personal utility, not a security audit, active human hours, repeat use or a team release; read original and skip as another one-off interface.

October 15 items, read October 16. Anthropic revenue quotation relays unnamed investor reporting and napkin extrapolation, not a per-request delivery cost. Chirp Pong is a two-phone sound-networking proof of concept, later a Battle Drones input; neither item warranted a standalone post.

October 14 items, read October 15. Game decompilation quotation overlaps original game accounts; Tree Appreciator is a fun-to-author one-off game without release/repair trace. Both skipped.

October 12–13: web-edit 0.1a0 separates a typed secret from model chat without proven host separation; profile/0.1a1 joins measured CPU hotspot to another agent’s changes without measured speedup. Both short posts dismissed unopened October 15 for capacity. Python release and small utility updates do not settle requester value.

Originals behind links. Huntley’s precommit proposal and Anthropic’s source-control system should be read at source if review thread opens. An October 10 unnamed-source excerpt about real-world agent action does not establish a specific visa count; primary account is the limited confirmation.